2 * linux/kernel/time/timekeeping.c
4 * Kernel timekeeping code and accessor functions
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
11 #include <linux/timekeeper_internal.h>
12 #include <linux/module.h>
13 #include <linux/interrupt.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
17 #include <linux/sched.h>
18 #include <linux/syscore_ops.h>
19 #include <linux/clocksource.h>
20 #include <linux/jiffies.h>
21 #include <linux/time.h>
22 #include <linux/tick.h>
23 #include <linux/stop_machine.h>
24 #include <linux/pvclock_gtod.h>
25 #include <linux/compiler.h>
27 #include "tick-internal.h"
28 #include "ntp_internal.h"
29 #include "timekeeping_internal.h"
31 #define TK_CLEAR_NTP (1 << 0)
32 #define TK_MIRROR (1 << 1)
33 #define TK_CLOCK_WAS_SET (1 << 2)
36 * The most important data for readout fits into a single 64 byte
41 struct timekeeper timekeeper;
42 } tk_core ____cacheline_aligned;
44 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
45 static struct timekeeper shadow_timekeeper;
48 * struct tk_fast - NMI safe timekeeper
49 * @seq: Sequence counter for protecting updates. The lowest bit
50 * is the index for the tk_read_base array
51 * @base: tk_read_base array. Access is indexed by the lowest bit of
54 * See @update_fast_timekeeper() below.
58 struct tk_read_base base[2];
61 static struct tk_fast tk_fast_mono ____cacheline_aligned;
62 static struct tk_fast tk_fast_raw ____cacheline_aligned;
64 /* flag for if timekeeping is suspended */
65 int __read_mostly timekeeping_suspended;
67 static inline void tk_normalize_xtime(struct timekeeper *tk)
69 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
70 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
75 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
79 ts.tv_sec = tk->xtime_sec;
80 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
84 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
86 tk->xtime_sec = ts->tv_sec;
87 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
90 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
92 tk->xtime_sec += ts->tv_sec;
93 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
94 tk_normalize_xtime(tk);
97 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
99 struct timespec64 tmp;
102 * Verify consistency of: offset_real = -wall_to_monotonic
103 * before modifying anything
105 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
106 -tk->wall_to_monotonic.tv_nsec);
107 WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
108 tk->wall_to_monotonic = wtm;
109 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
110 tk->offs_real = timespec64_to_ktime(tmp);
111 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
114 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
116 tk->offs_boot = ktime_add(tk->offs_boot, delta);
119 #ifdef CONFIG_DEBUG_TIMEKEEPING
120 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
122 static void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
125 cycle_t max_cycles = tk->tkr_mono.clock->max_cycles;
126 const char *name = tk->tkr_mono.clock->name;
128 if (offset > max_cycles) {
129 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
130 offset, name, max_cycles);
131 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
133 if (offset > (max_cycles >> 1)) {
134 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the the '%s' clock's 50%% safety margin (%lld)\n",
135 offset, name, max_cycles >> 1);
136 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
140 if (tk->underflow_seen) {
141 if (jiffies - tk->last_warning > WARNING_FREQ) {
142 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
143 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
144 printk_deferred(" Your kernel is probably still fine.\n");
145 tk->last_warning = jiffies;
147 tk->underflow_seen = 0;
150 if (tk->overflow_seen) {
151 if (jiffies - tk->last_warning > WARNING_FREQ) {
152 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
153 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
154 printk_deferred(" Your kernel is probably still fine.\n");
155 tk->last_warning = jiffies;
157 tk->overflow_seen = 0;
161 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
163 struct timekeeper *tk = &tk_core.timekeeper;
164 cycle_t now, last, mask, max, delta;
168 * Since we're called holding a seqlock, the data may shift
169 * under us while we're doing the calculation. This can cause
170 * false positives, since we'd note a problem but throw the
171 * results away. So nest another seqlock here to atomically
172 * grab the points we are checking with.
175 seq = read_seqcount_begin(&tk_core.seq);
176 now = tkr->read(tkr->clock);
177 last = tkr->cycle_last;
179 max = tkr->clock->max_cycles;
180 } while (read_seqcount_retry(&tk_core.seq, seq));
182 delta = clocksource_delta(now, last, mask);
185 * Try to catch underflows by checking if we are seeing small
186 * mask-relative negative values.
188 if (unlikely((~delta & mask) < (mask >> 3))) {
189 tk->underflow_seen = 1;
193 /* Cap delta value to the max_cycles values to avoid mult overflows */
194 if (unlikely(delta > max)) {
195 tk->overflow_seen = 1;
196 delta = tkr->clock->max_cycles;
202 static inline void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
205 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
207 cycle_t cycle_now, delta;
209 /* read clocksource */
210 cycle_now = tkr->read(tkr->clock);
212 /* calculate the delta since the last update_wall_time */
213 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
220 * tk_setup_internals - Set up internals to use clocksource clock.
222 * @tk: The target timekeeper to setup.
223 * @clock: Pointer to clocksource.
225 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
226 * pair and interval request.
228 * Unless you're the timekeeping code, you should not be using this!
230 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
233 u64 tmp, ntpinterval;
234 struct clocksource *old_clock;
236 old_clock = tk->tkr_mono.clock;
237 tk->tkr_mono.clock = clock;
238 tk->tkr_mono.read = clock->read;
239 tk->tkr_mono.mask = clock->mask;
240 tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
242 tk->tkr_raw.clock = clock;
243 tk->tkr_raw.read = clock->read;
244 tk->tkr_raw.mask = clock->mask;
245 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
247 /* Do the ns -> cycle conversion first, using original mult */
248 tmp = NTP_INTERVAL_LENGTH;
249 tmp <<= clock->shift;
251 tmp += clock->mult/2;
252 do_div(tmp, clock->mult);
256 interval = (cycle_t) tmp;
257 tk->cycle_interval = interval;
259 /* Go back from cycles -> shifted ns */
260 tk->xtime_interval = (u64) interval * clock->mult;
261 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
263 ((u64) interval * clock->mult) >> clock->shift;
265 /* if changing clocks, convert xtime_nsec shift units */
267 int shift_change = clock->shift - old_clock->shift;
268 if (shift_change < 0)
269 tk->tkr_mono.xtime_nsec >>= -shift_change;
271 tk->tkr_mono.xtime_nsec <<= shift_change;
273 tk->tkr_raw.xtime_nsec = 0;
275 tk->tkr_mono.shift = clock->shift;
276 tk->tkr_raw.shift = clock->shift;
279 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
280 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
283 * The timekeeper keeps its own mult values for the currently
284 * active clocksource. These value will be adjusted via NTP
285 * to counteract clock drifting.
287 tk->tkr_mono.mult = clock->mult;
288 tk->tkr_raw.mult = clock->mult;
289 tk->ntp_err_mult = 0;
292 /* Timekeeper helper functions. */
294 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
295 static u32 default_arch_gettimeoffset(void) { return 0; }
296 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
298 static inline u32 arch_gettimeoffset(void) { return 0; }
301 static inline u64 timekeeping_delta_to_ns(struct tk_read_base *tkr,
306 nsec = delta * tkr->mult + tkr->xtime_nsec;
309 /* If arch requires, add in get_arch_timeoffset() */
310 return nsec + arch_gettimeoffset();
313 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
317 delta = timekeeping_get_delta(tkr);
318 return timekeeping_delta_to_ns(tkr, delta);
321 static inline s64 timekeeping_cycles_to_ns(struct tk_read_base *tkr,
326 /* calculate the delta since the last update_wall_time */
327 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
328 return timekeeping_delta_to_ns(tkr, delta);
332 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
333 * @tkr: Timekeeping readout base from which we take the update
335 * We want to use this from any context including NMI and tracing /
336 * instrumenting the timekeeping code itself.
338 * Employ the latch technique; see @raw_write_seqcount_latch.
340 * So if a NMI hits the update of base[0] then it will use base[1]
341 * which is still consistent. In the worst case this can result is a
342 * slightly wrong timestamp (a few nanoseconds). See
343 * @ktime_get_mono_fast_ns.
345 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
347 struct tk_read_base *base = tkf->base;
349 /* Force readers off to base[1] */
350 raw_write_seqcount_latch(&tkf->seq);
353 memcpy(base, tkr, sizeof(*base));
355 /* Force readers back to base[0] */
356 raw_write_seqcount_latch(&tkf->seq);
359 memcpy(base + 1, base, sizeof(*base));
363 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
365 * This timestamp is not guaranteed to be monotonic across an update.
366 * The timestamp is calculated by:
368 * now = base_mono + clock_delta * slope
370 * So if the update lowers the slope, readers who are forced to the
371 * not yet updated second array are still using the old steeper slope.
380 * |12345678---> reader order
386 * So reader 6 will observe time going backwards versus reader 5.
388 * While other CPUs are likely to be able observe that, the only way
389 * for a CPU local observation is when an NMI hits in the middle of
390 * the update. Timestamps taken from that NMI context might be ahead
391 * of the following timestamps. Callers need to be aware of that and
394 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
396 struct tk_read_base *tkr;
401 seq = raw_read_seqcount_latch(&tkf->seq);
402 tkr = tkf->base + (seq & 0x01);
403 now = ktime_to_ns(tkr->base);
405 now += timekeeping_delta_to_ns(tkr,
407 tkr->read(tkr->clock),
410 } while (read_seqcount_retry(&tkf->seq, seq));
415 u64 ktime_get_mono_fast_ns(void)
417 return __ktime_get_fast_ns(&tk_fast_mono);
419 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
421 u64 ktime_get_raw_fast_ns(void)
423 return __ktime_get_fast_ns(&tk_fast_raw);
425 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
428 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
430 * To keep it NMI safe since we're accessing from tracing, we're not using a
431 * separate timekeeper with updates to monotonic clock and boot offset
432 * protected with seqlocks. This has the following minor side effects:
434 * (1) Its possible that a timestamp be taken after the boot offset is updated
435 * but before the timekeeper is updated. If this happens, the new boot offset
436 * is added to the old timekeeping making the clock appear to update slightly
439 * timekeeping_inject_sleeptime64()
440 * __timekeeping_inject_sleeptime(tk, delta);
442 * timekeeping_update(tk, TK_CLEAR_NTP...);
444 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
445 * partially updated. Since the tk->offs_boot update is a rare event, this
446 * should be a rare occurrence which postprocessing should be able to handle.
448 u64 notrace ktime_get_boot_fast_ns(void)
450 struct timekeeper *tk = &tk_core.timekeeper;
452 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
454 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
456 /* Suspend-time cycles value for halted fast timekeeper. */
457 static cycle_t cycles_at_suspend;
459 static cycle_t dummy_clock_read(struct clocksource *cs)
461 return cycles_at_suspend;
465 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
466 * @tk: Timekeeper to snapshot.
468 * It generally is unsafe to access the clocksource after timekeeping has been
469 * suspended, so take a snapshot of the readout base of @tk and use it as the
470 * fast timekeeper's readout base while suspended. It will return the same
471 * number of cycles every time until timekeeping is resumed at which time the
472 * proper readout base for the fast timekeeper will be restored automatically.
474 static void halt_fast_timekeeper(struct timekeeper *tk)
476 static struct tk_read_base tkr_dummy;
477 struct tk_read_base *tkr = &tk->tkr_mono;
479 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
480 cycles_at_suspend = tkr->read(tkr->clock);
481 tkr_dummy.read = dummy_clock_read;
482 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
485 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
486 tkr_dummy.read = dummy_clock_read;
487 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
490 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
492 static inline void update_vsyscall(struct timekeeper *tk)
494 struct timespec xt, wm;
496 xt = timespec64_to_timespec(tk_xtime(tk));
497 wm = timespec64_to_timespec(tk->wall_to_monotonic);
498 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
499 tk->tkr_mono.cycle_last);
502 static inline void old_vsyscall_fixup(struct timekeeper *tk)
507 * Store only full nanoseconds into xtime_nsec after rounding
508 * it up and add the remainder to the error difference.
509 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
510 * by truncating the remainder in vsyscalls. However, it causes
511 * additional work to be done in timekeeping_adjust(). Once
512 * the vsyscall implementations are converted to use xtime_nsec
513 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
514 * users are removed, this can be killed.
516 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
517 tk->tkr_mono.xtime_nsec -= remainder;
518 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
519 tk->ntp_error += remainder << tk->ntp_error_shift;
520 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
523 #define old_vsyscall_fixup(tk)
526 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
528 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
530 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
534 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
536 int pvclock_gtod_register_notifier(struct notifier_block *nb)
538 struct timekeeper *tk = &tk_core.timekeeper;
542 raw_spin_lock_irqsave(&timekeeper_lock, flags);
543 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
544 update_pvclock_gtod(tk, true);
545 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
549 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
552 * pvclock_gtod_unregister_notifier - unregister a pvclock
553 * timedata update listener
555 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
560 raw_spin_lock_irqsave(&timekeeper_lock, flags);
561 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
562 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
566 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
569 * tk_update_leap_state - helper to update the next_leap_ktime
571 static inline void tk_update_leap_state(struct timekeeper *tk)
573 tk->next_leap_ktime = ntp_get_next_leap();
574 if (tk->next_leap_ktime.tv64 != KTIME_MAX)
575 /* Convert to monotonic time */
576 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
580 * Update the ktime_t based scalar nsec members of the timekeeper
582 static inline void tk_update_ktime_data(struct timekeeper *tk)
588 * The xtime based monotonic readout is:
589 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
590 * The ktime based monotonic readout is:
591 * nsec = base_mono + now();
592 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
594 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
595 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
596 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
598 /* Update the monotonic raw base */
599 tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
602 * The sum of the nanoseconds portions of xtime and
603 * wall_to_monotonic can be greater/equal one second. Take
604 * this into account before updating tk->ktime_sec.
606 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
607 if (nsec >= NSEC_PER_SEC)
609 tk->ktime_sec = seconds;
612 /* must hold timekeeper_lock */
613 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
615 if (action & TK_CLEAR_NTP) {
620 tk_update_leap_state(tk);
621 tk_update_ktime_data(tk);
624 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
626 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
627 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
629 if (action & TK_CLOCK_WAS_SET)
630 tk->clock_was_set_seq++;
632 * The mirroring of the data to the shadow-timekeeper needs
633 * to happen last here to ensure we don't over-write the
634 * timekeeper structure on the next update with stale data
636 if (action & TK_MIRROR)
637 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
638 sizeof(tk_core.timekeeper));
642 * timekeeping_forward_now - update clock to the current time
644 * Forward the current clock to update its state since the last call to
645 * update_wall_time(). This is useful before significant clock changes,
646 * as it avoids having to deal with this time offset explicitly.
648 static void timekeeping_forward_now(struct timekeeper *tk)
650 struct clocksource *clock = tk->tkr_mono.clock;
651 cycle_t cycle_now, delta;
654 cycle_now = tk->tkr_mono.read(clock);
655 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
656 tk->tkr_mono.cycle_last = cycle_now;
657 tk->tkr_raw.cycle_last = cycle_now;
659 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
661 /* If arch requires, add in get_arch_timeoffset() */
662 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
664 tk_normalize_xtime(tk);
666 nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
667 timespec64_add_ns(&tk->raw_time, nsec);
671 * __getnstimeofday64 - Returns the time of day in a timespec64.
672 * @ts: pointer to the timespec to be set
674 * Updates the time of day in the timespec.
675 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
677 int __getnstimeofday64(struct timespec64 *ts)
679 struct timekeeper *tk = &tk_core.timekeeper;
684 seq = read_seqcount_begin(&tk_core.seq);
686 ts->tv_sec = tk->xtime_sec;
687 nsecs = timekeeping_get_ns(&tk->tkr_mono);
689 } while (read_seqcount_retry(&tk_core.seq, seq));
692 timespec64_add_ns(ts, nsecs);
695 * Do not bail out early, in case there were callers still using
696 * the value, even in the face of the WARN_ON.
698 if (unlikely(timekeeping_suspended))
702 EXPORT_SYMBOL(__getnstimeofday64);
705 * getnstimeofday64 - Returns the time of day in a timespec64.
706 * @ts: pointer to the timespec64 to be set
708 * Returns the time of day in a timespec64 (WARN if suspended).
710 void getnstimeofday64(struct timespec64 *ts)
712 WARN_ON(__getnstimeofday64(ts));
714 EXPORT_SYMBOL(getnstimeofday64);
716 ktime_t ktime_get(void)
718 struct timekeeper *tk = &tk_core.timekeeper;
723 WARN_ON(timekeeping_suspended);
726 seq = read_seqcount_begin(&tk_core.seq);
727 base = tk->tkr_mono.base;
728 nsecs = timekeeping_get_ns(&tk->tkr_mono);
730 } while (read_seqcount_retry(&tk_core.seq, seq));
732 return ktime_add_ns(base, nsecs);
734 EXPORT_SYMBOL_GPL(ktime_get);
736 u32 ktime_get_resolution_ns(void)
738 struct timekeeper *tk = &tk_core.timekeeper;
742 WARN_ON(timekeeping_suspended);
745 seq = read_seqcount_begin(&tk_core.seq);
746 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
747 } while (read_seqcount_retry(&tk_core.seq, seq));
751 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
753 static ktime_t *offsets[TK_OFFS_MAX] = {
754 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
755 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
756 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
759 ktime_t ktime_get_with_offset(enum tk_offsets offs)
761 struct timekeeper *tk = &tk_core.timekeeper;
763 ktime_t base, *offset = offsets[offs];
766 WARN_ON(timekeeping_suspended);
769 seq = read_seqcount_begin(&tk_core.seq);
770 base = ktime_add(tk->tkr_mono.base, *offset);
771 nsecs = timekeeping_get_ns(&tk->tkr_mono);
773 } while (read_seqcount_retry(&tk_core.seq, seq));
775 return ktime_add_ns(base, nsecs);
778 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
781 * ktime_mono_to_any() - convert mononotic time to any other time
782 * @tmono: time to convert.
783 * @offs: which offset to use
785 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
787 ktime_t *offset = offsets[offs];
792 seq = read_seqcount_begin(&tk_core.seq);
793 tconv = ktime_add(tmono, *offset);
794 } while (read_seqcount_retry(&tk_core.seq, seq));
798 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
801 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
803 ktime_t ktime_get_raw(void)
805 struct timekeeper *tk = &tk_core.timekeeper;
811 seq = read_seqcount_begin(&tk_core.seq);
812 base = tk->tkr_raw.base;
813 nsecs = timekeeping_get_ns(&tk->tkr_raw);
815 } while (read_seqcount_retry(&tk_core.seq, seq));
817 return ktime_add_ns(base, nsecs);
819 EXPORT_SYMBOL_GPL(ktime_get_raw);
822 * ktime_get_ts64 - get the monotonic clock in timespec64 format
823 * @ts: pointer to timespec variable
825 * The function calculates the monotonic clock from the realtime
826 * clock and the wall_to_monotonic offset and stores the result
827 * in normalized timespec64 format in the variable pointed to by @ts.
829 void ktime_get_ts64(struct timespec64 *ts)
831 struct timekeeper *tk = &tk_core.timekeeper;
832 struct timespec64 tomono;
836 WARN_ON(timekeeping_suspended);
839 seq = read_seqcount_begin(&tk_core.seq);
840 ts->tv_sec = tk->xtime_sec;
841 nsec = timekeeping_get_ns(&tk->tkr_mono);
842 tomono = tk->wall_to_monotonic;
844 } while (read_seqcount_retry(&tk_core.seq, seq));
846 ts->tv_sec += tomono.tv_sec;
848 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
850 EXPORT_SYMBOL_GPL(ktime_get_ts64);
853 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
855 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
856 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
857 * works on both 32 and 64 bit systems. On 32 bit systems the readout
858 * covers ~136 years of uptime which should be enough to prevent
859 * premature wrap arounds.
861 time64_t ktime_get_seconds(void)
863 struct timekeeper *tk = &tk_core.timekeeper;
865 WARN_ON(timekeeping_suspended);
866 return tk->ktime_sec;
868 EXPORT_SYMBOL_GPL(ktime_get_seconds);
871 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
873 * Returns the wall clock seconds since 1970. This replaces the
874 * get_seconds() interface which is not y2038 safe on 32bit systems.
876 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
877 * 32bit systems the access must be protected with the sequence
878 * counter to provide "atomic" access to the 64bit tk->xtime_sec
881 time64_t ktime_get_real_seconds(void)
883 struct timekeeper *tk = &tk_core.timekeeper;
887 if (IS_ENABLED(CONFIG_64BIT))
888 return tk->xtime_sec;
891 seq = read_seqcount_begin(&tk_core.seq);
892 seconds = tk->xtime_sec;
894 } while (read_seqcount_retry(&tk_core.seq, seq));
898 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
900 #ifdef CONFIG_NTP_PPS
903 * ktime_get_raw_and_real_ts64 - get day and raw monotonic time in timespec format
904 * @ts_raw: pointer to the timespec to be set to raw monotonic time
905 * @ts_real: pointer to the timespec to be set to the time of day
907 * This function reads both the time of day and raw monotonic time at the
908 * same time atomically and stores the resulting timestamps in timespec
911 void ktime_get_raw_and_real_ts64(struct timespec64 *ts_raw, struct timespec64 *ts_real)
913 struct timekeeper *tk = &tk_core.timekeeper;
915 s64 nsecs_raw, nsecs_real;
917 WARN_ON_ONCE(timekeeping_suspended);
920 seq = read_seqcount_begin(&tk_core.seq);
922 *ts_raw = tk->raw_time;
923 ts_real->tv_sec = tk->xtime_sec;
924 ts_real->tv_nsec = 0;
926 nsecs_raw = timekeeping_get_ns(&tk->tkr_raw);
927 nsecs_real = timekeeping_get_ns(&tk->tkr_mono);
929 } while (read_seqcount_retry(&tk_core.seq, seq));
931 timespec64_add_ns(ts_raw, nsecs_raw);
932 timespec64_add_ns(ts_real, nsecs_real);
934 EXPORT_SYMBOL(ktime_get_raw_and_real_ts64);
936 #endif /* CONFIG_NTP_PPS */
939 * do_gettimeofday - Returns the time of day in a timeval
940 * @tv: pointer to the timeval to be set
942 * NOTE: Users should be converted to using getnstimeofday()
944 void do_gettimeofday(struct timeval *tv)
946 struct timespec64 now;
948 getnstimeofday64(&now);
949 tv->tv_sec = now.tv_sec;
950 tv->tv_usec = now.tv_nsec/1000;
952 EXPORT_SYMBOL(do_gettimeofday);
955 * do_settimeofday64 - Sets the time of day.
956 * @ts: pointer to the timespec64 variable containing the new time
958 * Sets the time of day to the new time and update NTP and notify hrtimers
960 int do_settimeofday64(const struct timespec64 *ts)
962 struct timekeeper *tk = &tk_core.timekeeper;
963 struct timespec64 ts_delta, xt;
967 if (!timespec64_valid_strict(ts))
970 raw_spin_lock_irqsave(&timekeeper_lock, flags);
971 write_seqcount_begin(&tk_core.seq);
973 timekeeping_forward_now(tk);
976 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
977 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
979 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
984 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
986 tk_set_xtime(tk, ts);
988 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
990 write_seqcount_end(&tk_core.seq);
991 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
993 /* signal hrtimers about time change */
998 EXPORT_SYMBOL(do_settimeofday64);
1001 * timekeeping_inject_offset - Adds or subtracts from the current time.
1002 * @tv: pointer to the timespec variable containing the offset
1004 * Adds or subtracts an offset value from the current time.
1006 int timekeeping_inject_offset(struct timespec *ts)
1008 struct timekeeper *tk = &tk_core.timekeeper;
1009 unsigned long flags;
1010 struct timespec64 ts64, tmp;
1013 if (!timespec_inject_offset_valid(ts))
1016 ts64 = timespec_to_timespec64(*ts);
1018 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1019 write_seqcount_begin(&tk_core.seq);
1021 timekeeping_forward_now(tk);
1023 /* Make sure the proposed value is valid */
1024 tmp = timespec64_add(tk_xtime(tk), ts64);
1025 if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
1026 !timespec64_valid_strict(&tmp)) {
1031 tk_xtime_add(tk, &ts64);
1032 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
1034 error: /* even if we error out, we forwarded the time, so call update */
1035 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1037 write_seqcount_end(&tk_core.seq);
1038 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1040 /* signal hrtimers about time change */
1045 EXPORT_SYMBOL(timekeeping_inject_offset);
1049 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1052 s32 timekeeping_get_tai_offset(void)
1054 struct timekeeper *tk = &tk_core.timekeeper;
1059 seq = read_seqcount_begin(&tk_core.seq);
1060 ret = tk->tai_offset;
1061 } while (read_seqcount_retry(&tk_core.seq, seq));
1067 * __timekeeping_set_tai_offset - Lock free worker function
1070 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1072 tk->tai_offset = tai_offset;
1073 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1077 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1080 void timekeeping_set_tai_offset(s32 tai_offset)
1082 struct timekeeper *tk = &tk_core.timekeeper;
1083 unsigned long flags;
1085 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1086 write_seqcount_begin(&tk_core.seq);
1087 __timekeeping_set_tai_offset(tk, tai_offset);
1088 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1089 write_seqcount_end(&tk_core.seq);
1090 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1095 * change_clocksource - Swaps clocksources if a new one is available
1097 * Accumulates current time interval and initializes new clocksource
1099 static int change_clocksource(void *data)
1101 struct timekeeper *tk = &tk_core.timekeeper;
1102 struct clocksource *new, *old;
1103 unsigned long flags;
1105 new = (struct clocksource *) data;
1107 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1108 write_seqcount_begin(&tk_core.seq);
1110 timekeeping_forward_now(tk);
1112 * If the cs is in module, get a module reference. Succeeds
1113 * for built-in code (owner == NULL) as well.
1115 if (try_module_get(new->owner)) {
1116 if (!new->enable || new->enable(new) == 0) {
1117 old = tk->tkr_mono.clock;
1118 tk_setup_internals(tk, new);
1121 module_put(old->owner);
1123 module_put(new->owner);
1126 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1128 write_seqcount_end(&tk_core.seq);
1129 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1135 * timekeeping_notify - Install a new clock source
1136 * @clock: pointer to the clock source
1138 * This function is called from clocksource.c after a new, better clock
1139 * source has been registered. The caller holds the clocksource_mutex.
1141 int timekeeping_notify(struct clocksource *clock)
1143 struct timekeeper *tk = &tk_core.timekeeper;
1145 if (tk->tkr_mono.clock == clock)
1147 stop_machine(change_clocksource, clock, NULL);
1148 tick_clock_notify();
1149 return tk->tkr_mono.clock == clock ? 0 : -1;
1153 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1154 * @ts: pointer to the timespec64 to be set
1156 * Returns the raw monotonic time (completely un-modified by ntp)
1158 void getrawmonotonic64(struct timespec64 *ts)
1160 struct timekeeper *tk = &tk_core.timekeeper;
1161 struct timespec64 ts64;
1166 seq = read_seqcount_begin(&tk_core.seq);
1167 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1168 ts64 = tk->raw_time;
1170 } while (read_seqcount_retry(&tk_core.seq, seq));
1172 timespec64_add_ns(&ts64, nsecs);
1175 EXPORT_SYMBOL(getrawmonotonic64);
1179 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1181 int timekeeping_valid_for_hres(void)
1183 struct timekeeper *tk = &tk_core.timekeeper;
1188 seq = read_seqcount_begin(&tk_core.seq);
1190 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1192 } while (read_seqcount_retry(&tk_core.seq, seq));
1198 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1200 u64 timekeeping_max_deferment(void)
1202 struct timekeeper *tk = &tk_core.timekeeper;
1207 seq = read_seqcount_begin(&tk_core.seq);
1209 ret = tk->tkr_mono.clock->max_idle_ns;
1211 } while (read_seqcount_retry(&tk_core.seq, seq));
1217 * read_persistent_clock - Return time from the persistent clock.
1219 * Weak dummy function for arches that do not yet support it.
1220 * Reads the time from the battery backed persistent clock.
1221 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1223 * XXX - Do be sure to remove it once all arches implement it.
1225 void __weak read_persistent_clock(struct timespec *ts)
1231 void __weak read_persistent_clock64(struct timespec64 *ts64)
1235 read_persistent_clock(&ts);
1236 *ts64 = timespec_to_timespec64(ts);
1240 * read_boot_clock64 - Return time of the system start.
1242 * Weak dummy function for arches that do not yet support it.
1243 * Function to read the exact time the system has been started.
1244 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1246 * XXX - Do be sure to remove it once all arches implement it.
1248 void __weak read_boot_clock64(struct timespec64 *ts)
1254 /* Flag for if timekeeping_resume() has injected sleeptime */
1255 static bool sleeptime_injected;
1257 /* Flag for if there is a persistent clock on this platform */
1258 static bool persistent_clock_exists;
1261 * timekeeping_init - Initializes the clocksource and common timekeeping values
1263 void __init timekeeping_init(void)
1265 struct timekeeper *tk = &tk_core.timekeeper;
1266 struct clocksource *clock;
1267 unsigned long flags;
1268 struct timespec64 now, boot, tmp;
1270 read_persistent_clock64(&now);
1271 if (!timespec64_valid_strict(&now)) {
1272 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1273 " Check your CMOS/BIOS settings.\n");
1276 } else if (now.tv_sec || now.tv_nsec)
1277 persistent_clock_exists = true;
1279 read_boot_clock64(&boot);
1280 if (!timespec64_valid_strict(&boot)) {
1281 pr_warn("WARNING: Boot clock returned invalid value!\n"
1282 " Check your CMOS/BIOS settings.\n");
1287 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1288 write_seqcount_begin(&tk_core.seq);
1291 clock = clocksource_default_clock();
1293 clock->enable(clock);
1294 tk_setup_internals(tk, clock);
1296 tk_set_xtime(tk, &now);
1297 tk->raw_time.tv_sec = 0;
1298 tk->raw_time.tv_nsec = 0;
1299 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1300 boot = tk_xtime(tk);
1302 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1303 tk_set_wall_to_mono(tk, tmp);
1305 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1307 write_seqcount_end(&tk_core.seq);
1308 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1311 /* time in seconds when suspend began for persistent clock */
1312 static struct timespec64 timekeeping_suspend_time;
1315 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1316 * @delta: pointer to a timespec delta value
1318 * Takes a timespec offset measuring a suspend interval and properly
1319 * adds the sleep offset to the timekeeping variables.
1321 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1322 struct timespec64 *delta)
1324 if (!timespec64_valid_strict(delta)) {
1325 printk_deferred(KERN_WARNING
1326 "__timekeeping_inject_sleeptime: Invalid "
1327 "sleep delta value!\n");
1330 tk_xtime_add(tk, delta);
1331 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1332 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1333 tk_debug_account_sleep_time(delta);
1336 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1338 * We have three kinds of time sources to use for sleep time
1339 * injection, the preference order is:
1340 * 1) non-stop clocksource
1341 * 2) persistent clock (ie: RTC accessible when irqs are off)
1344 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1345 * If system has neither 1) nor 2), 3) will be used finally.
1348 * If timekeeping has injected sleeptime via either 1) or 2),
1349 * 3) becomes needless, so in this case we don't need to call
1350 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1353 bool timekeeping_rtc_skipresume(void)
1355 return sleeptime_injected;
1359 * 1) can be determined whether to use or not only when doing
1360 * timekeeping_resume() which is invoked after rtc_suspend(),
1361 * so we can't skip rtc_suspend() surely if system has 1).
1363 * But if system has 2), 2) will definitely be used, so in this
1364 * case we don't need to call rtc_suspend(), and this is what
1365 * timekeeping_rtc_skipsuspend() means.
1367 bool timekeeping_rtc_skipsuspend(void)
1369 return persistent_clock_exists;
1373 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1374 * @delta: pointer to a timespec64 delta value
1376 * This hook is for architectures that cannot support read_persistent_clock64
1377 * because their RTC/persistent clock is only accessible when irqs are enabled.
1378 * and also don't have an effective nonstop clocksource.
1380 * This function should only be called by rtc_resume(), and allows
1381 * a suspend offset to be injected into the timekeeping values.
1383 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1385 struct timekeeper *tk = &tk_core.timekeeper;
1386 unsigned long flags;
1388 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1389 write_seqcount_begin(&tk_core.seq);
1391 timekeeping_forward_now(tk);
1393 __timekeeping_inject_sleeptime(tk, delta);
1395 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1397 write_seqcount_end(&tk_core.seq);
1398 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1400 /* signal hrtimers about time change */
1406 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1408 void timekeeping_resume(void)
1410 struct timekeeper *tk = &tk_core.timekeeper;
1411 struct clocksource *clock = tk->tkr_mono.clock;
1412 unsigned long flags;
1413 struct timespec64 ts_new, ts_delta;
1414 cycle_t cycle_now, cycle_delta;
1416 sleeptime_injected = false;
1417 read_persistent_clock64(&ts_new);
1419 clockevents_resume();
1420 clocksource_resume();
1422 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1423 write_seqcount_begin(&tk_core.seq);
1426 * After system resumes, we need to calculate the suspended time and
1427 * compensate it for the OS time. There are 3 sources that could be
1428 * used: Nonstop clocksource during suspend, persistent clock and rtc
1431 * One specific platform may have 1 or 2 or all of them, and the
1432 * preference will be:
1433 * suspend-nonstop clocksource -> persistent clock -> rtc
1434 * The less preferred source will only be tried if there is no better
1435 * usable source. The rtc part is handled separately in rtc core code.
1437 cycle_now = tk->tkr_mono.read(clock);
1438 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1439 cycle_now > tk->tkr_mono.cycle_last) {
1440 u64 num, max = ULLONG_MAX;
1441 u32 mult = clock->mult;
1442 u32 shift = clock->shift;
1445 cycle_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1449 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1450 * suspended time is too long. In that case we need do the
1451 * 64 bits math carefully
1454 if (cycle_delta > max) {
1455 num = div64_u64(cycle_delta, max);
1456 nsec = (((u64) max * mult) >> shift) * num;
1457 cycle_delta -= num * max;
1459 nsec += ((u64) cycle_delta * mult) >> shift;
1461 ts_delta = ns_to_timespec64(nsec);
1462 sleeptime_injected = true;
1463 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1464 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1465 sleeptime_injected = true;
1468 if (sleeptime_injected)
1469 __timekeeping_inject_sleeptime(tk, &ts_delta);
1471 /* Re-base the last cycle value */
1472 tk->tkr_mono.cycle_last = cycle_now;
1473 tk->tkr_raw.cycle_last = cycle_now;
1476 timekeeping_suspended = 0;
1477 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1478 write_seqcount_end(&tk_core.seq);
1479 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1481 touch_softlockup_watchdog();
1487 int timekeeping_suspend(void)
1489 struct timekeeper *tk = &tk_core.timekeeper;
1490 unsigned long flags;
1491 struct timespec64 delta, delta_delta;
1492 static struct timespec64 old_delta;
1494 read_persistent_clock64(&timekeeping_suspend_time);
1497 * On some systems the persistent_clock can not be detected at
1498 * timekeeping_init by its return value, so if we see a valid
1499 * value returned, update the persistent_clock_exists flag.
1501 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1502 persistent_clock_exists = true;
1504 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1505 write_seqcount_begin(&tk_core.seq);
1506 timekeeping_forward_now(tk);
1507 timekeeping_suspended = 1;
1509 if (persistent_clock_exists) {
1511 * To avoid drift caused by repeated suspend/resumes,
1512 * which each can add ~1 second drift error,
1513 * try to compensate so the difference in system time
1514 * and persistent_clock time stays close to constant.
1516 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1517 delta_delta = timespec64_sub(delta, old_delta);
1518 if (abs(delta_delta.tv_sec) >= 2) {
1520 * if delta_delta is too large, assume time correction
1521 * has occurred and set old_delta to the current delta.
1525 /* Otherwise try to adjust old_system to compensate */
1526 timekeeping_suspend_time =
1527 timespec64_add(timekeeping_suspend_time, delta_delta);
1531 timekeeping_update(tk, TK_MIRROR);
1532 halt_fast_timekeeper(tk);
1533 write_seqcount_end(&tk_core.seq);
1534 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1537 clocksource_suspend();
1538 clockevents_suspend();
1543 /* sysfs resume/suspend bits for timekeeping */
1544 static struct syscore_ops timekeeping_syscore_ops = {
1545 .resume = timekeeping_resume,
1546 .suspend = timekeeping_suspend,
1549 static int __init timekeeping_init_ops(void)
1551 register_syscore_ops(&timekeeping_syscore_ops);
1554 device_initcall(timekeeping_init_ops);
1557 * Apply a multiplier adjustment to the timekeeper
1559 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1564 s64 interval = tk->cycle_interval;
1568 mult_adj = -mult_adj;
1569 interval = -interval;
1572 mult_adj <<= adj_scale;
1573 interval <<= adj_scale;
1574 offset <<= adj_scale;
1577 * So the following can be confusing.
1579 * To keep things simple, lets assume mult_adj == 1 for now.
1581 * When mult_adj != 1, remember that the interval and offset values
1582 * have been appropriately scaled so the math is the same.
1584 * The basic idea here is that we're increasing the multiplier
1585 * by one, this causes the xtime_interval to be incremented by
1586 * one cycle_interval. This is because:
1587 * xtime_interval = cycle_interval * mult
1588 * So if mult is being incremented by one:
1589 * xtime_interval = cycle_interval * (mult + 1)
1591 * xtime_interval = (cycle_interval * mult) + cycle_interval
1592 * Which can be shortened to:
1593 * xtime_interval += cycle_interval
1595 * So offset stores the non-accumulated cycles. Thus the current
1596 * time (in shifted nanoseconds) is:
1597 * now = (offset * adj) + xtime_nsec
1598 * Now, even though we're adjusting the clock frequency, we have
1599 * to keep time consistent. In other words, we can't jump back
1600 * in time, and we also want to avoid jumping forward in time.
1602 * So given the same offset value, we need the time to be the same
1603 * both before and after the freq adjustment.
1604 * now = (offset * adj_1) + xtime_nsec_1
1605 * now = (offset * adj_2) + xtime_nsec_2
1607 * (offset * adj_1) + xtime_nsec_1 =
1608 * (offset * adj_2) + xtime_nsec_2
1612 * (offset * adj_1) + xtime_nsec_1 =
1613 * (offset * (adj_1+1)) + xtime_nsec_2
1614 * (offset * adj_1) + xtime_nsec_1 =
1615 * (offset * adj_1) + offset + xtime_nsec_2
1616 * Canceling the sides:
1617 * xtime_nsec_1 = offset + xtime_nsec_2
1619 * xtime_nsec_2 = xtime_nsec_1 - offset
1620 * Which simplfies to:
1621 * xtime_nsec -= offset
1623 * XXX - TODO: Doc ntp_error calculation.
1625 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1626 /* NTP adjustment caused clocksource mult overflow */
1631 tk->tkr_mono.mult += mult_adj;
1632 tk->xtime_interval += interval;
1633 tk->tkr_mono.xtime_nsec -= offset;
1634 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1638 * Calculate the multiplier adjustment needed to match the frequency
1641 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1644 s64 interval = tk->cycle_interval;
1645 s64 xinterval = tk->xtime_interval;
1650 /* Remove any current error adj from freq calculation */
1651 if (tk->ntp_err_mult)
1652 xinterval -= tk->cycle_interval;
1654 tk->ntp_tick = ntp_tick_length();
1656 /* Calculate current error per tick */
1657 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1658 tick_error -= (xinterval + tk->xtime_remainder);
1660 /* Don't worry about correcting it if its small */
1661 if (likely((tick_error >= 0) && (tick_error <= interval)))
1664 /* preserve the direction of correction */
1665 negative = (tick_error < 0);
1667 /* Sort out the magnitude of the correction */
1668 tick_error = abs(tick_error);
1669 for (adj = 0; tick_error > interval; adj++)
1672 /* scale the corrections */
1673 timekeeping_apply_adjustment(tk, offset, negative, adj);
1677 * Adjust the timekeeper's multiplier to the correct frequency
1678 * and also to reduce the accumulated error value.
1680 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1682 /* Correct for the current frequency error */
1683 timekeeping_freqadjust(tk, offset);
1685 /* Next make a small adjustment to fix any cumulative error */
1686 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1687 tk->ntp_err_mult = 1;
1688 timekeeping_apply_adjustment(tk, offset, 0, 0);
1689 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1690 /* Undo any existing error adjustment */
1691 timekeeping_apply_adjustment(tk, offset, 1, 0);
1692 tk->ntp_err_mult = 0;
1695 if (unlikely(tk->tkr_mono.clock->maxadj &&
1696 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1697 > tk->tkr_mono.clock->maxadj))) {
1698 printk_once(KERN_WARNING
1699 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1700 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1701 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1705 * It may be possible that when we entered this function, xtime_nsec
1706 * was very small. Further, if we're slightly speeding the clocksource
1707 * in the code above, its possible the required corrective factor to
1708 * xtime_nsec could cause it to underflow.
1710 * Now, since we already accumulated the second, cannot simply roll
1711 * the accumulated second back, since the NTP subsystem has been
1712 * notified via second_overflow. So instead we push xtime_nsec forward
1713 * by the amount we underflowed, and add that amount into the error.
1715 * We'll correct this error next time through this function, when
1716 * xtime_nsec is not as small.
1718 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1719 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1720 tk->tkr_mono.xtime_nsec = 0;
1721 tk->ntp_error += neg << tk->ntp_error_shift;
1726 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1728 * Helper function that accumulates the nsecs greater than a second
1729 * from the xtime_nsec field to the xtime_secs field.
1730 * It also calls into the NTP code to handle leapsecond processing.
1733 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1735 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1736 unsigned int clock_set = 0;
1738 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1741 tk->tkr_mono.xtime_nsec -= nsecps;
1744 /* Figure out if its a leap sec and apply if needed */
1745 leap = second_overflow(tk->xtime_sec);
1746 if (unlikely(leap)) {
1747 struct timespec64 ts;
1749 tk->xtime_sec += leap;
1753 tk_set_wall_to_mono(tk,
1754 timespec64_sub(tk->wall_to_monotonic, ts));
1756 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1758 clock_set = TK_CLOCK_WAS_SET;
1765 * logarithmic_accumulation - shifted accumulation of cycles
1767 * This functions accumulates a shifted interval of cycles into
1768 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1771 * Returns the unconsumed cycles.
1773 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1775 unsigned int *clock_set)
1777 cycle_t interval = tk->cycle_interval << shift;
1780 /* If the offset is smaller than a shifted interval, do nothing */
1781 if (offset < interval)
1784 /* Accumulate one shifted interval */
1786 tk->tkr_mono.cycle_last += interval;
1787 tk->tkr_raw.cycle_last += interval;
1789 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
1790 *clock_set |= accumulate_nsecs_to_secs(tk);
1792 /* Accumulate raw time */
1793 raw_nsecs = (u64)tk->raw_interval << shift;
1794 raw_nsecs += tk->raw_time.tv_nsec;
1795 if (raw_nsecs >= NSEC_PER_SEC) {
1796 u64 raw_secs = raw_nsecs;
1797 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1798 tk->raw_time.tv_sec += raw_secs;
1800 tk->raw_time.tv_nsec = raw_nsecs;
1802 /* Accumulate error between NTP and clock interval */
1803 tk->ntp_error += tk->ntp_tick << shift;
1804 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1805 (tk->ntp_error_shift + shift);
1811 * update_wall_time - Uses the current clocksource to increment the wall time
1814 void update_wall_time(void)
1816 struct timekeeper *real_tk = &tk_core.timekeeper;
1817 struct timekeeper *tk = &shadow_timekeeper;
1819 int shift = 0, maxshift;
1820 unsigned int clock_set = 0;
1821 unsigned long flags;
1823 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1825 /* Make sure we're fully resumed: */
1826 if (unlikely(timekeeping_suspended))
1829 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1830 offset = real_tk->cycle_interval;
1832 offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
1833 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
1836 /* Check if there's really nothing to do */
1837 if (offset < real_tk->cycle_interval)
1840 /* Do some additional sanity checking */
1841 timekeeping_check_update(real_tk, offset);
1844 * With NO_HZ we may have to accumulate many cycle_intervals
1845 * (think "ticks") worth of time at once. To do this efficiently,
1846 * we calculate the largest doubling multiple of cycle_intervals
1847 * that is smaller than the offset. We then accumulate that
1848 * chunk in one go, and then try to consume the next smaller
1851 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1852 shift = max(0, shift);
1853 /* Bound shift to one less than what overflows tick_length */
1854 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1855 shift = min(shift, maxshift);
1856 while (offset >= tk->cycle_interval) {
1857 offset = logarithmic_accumulation(tk, offset, shift,
1859 if (offset < tk->cycle_interval<<shift)
1863 /* correct the clock when NTP error is too big */
1864 timekeeping_adjust(tk, offset);
1867 * XXX This can be killed once everyone converts
1868 * to the new update_vsyscall.
1870 old_vsyscall_fixup(tk);
1873 * Finally, make sure that after the rounding
1874 * xtime_nsec isn't larger than NSEC_PER_SEC
1876 clock_set |= accumulate_nsecs_to_secs(tk);
1878 write_seqcount_begin(&tk_core.seq);
1880 * Update the real timekeeper.
1882 * We could avoid this memcpy by switching pointers, but that
1883 * requires changes to all other timekeeper usage sites as
1884 * well, i.e. move the timekeeper pointer getter into the
1885 * spinlocked/seqcount protected sections. And we trade this
1886 * memcpy under the tk_core.seq against one before we start
1889 timekeeping_update(tk, clock_set);
1890 memcpy(real_tk, tk, sizeof(*tk));
1891 /* The memcpy must come last. Do not put anything here! */
1892 write_seqcount_end(&tk_core.seq);
1894 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1896 /* Have to call _delayed version, since in irq context*/
1897 clock_was_set_delayed();
1901 * getboottime64 - Return the real time of system boot.
1902 * @ts: pointer to the timespec64 to be set
1904 * Returns the wall-time of boot in a timespec64.
1906 * This is based on the wall_to_monotonic offset and the total suspend
1907 * time. Calls to settimeofday will affect the value returned (which
1908 * basically means that however wrong your real time clock is at boot time,
1909 * you get the right time here).
1911 void getboottime64(struct timespec64 *ts)
1913 struct timekeeper *tk = &tk_core.timekeeper;
1914 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1916 *ts = ktime_to_timespec64(t);
1918 EXPORT_SYMBOL_GPL(getboottime64);
1920 unsigned long get_seconds(void)
1922 struct timekeeper *tk = &tk_core.timekeeper;
1924 return tk->xtime_sec;
1926 EXPORT_SYMBOL(get_seconds);
1928 struct timespec __current_kernel_time(void)
1930 struct timekeeper *tk = &tk_core.timekeeper;
1932 return timespec64_to_timespec(tk_xtime(tk));
1935 struct timespec64 current_kernel_time64(void)
1937 struct timekeeper *tk = &tk_core.timekeeper;
1938 struct timespec64 now;
1942 seq = read_seqcount_begin(&tk_core.seq);
1945 } while (read_seqcount_retry(&tk_core.seq, seq));
1949 EXPORT_SYMBOL(current_kernel_time64);
1951 struct timespec64 get_monotonic_coarse64(void)
1953 struct timekeeper *tk = &tk_core.timekeeper;
1954 struct timespec64 now, mono;
1958 seq = read_seqcount_begin(&tk_core.seq);
1961 mono = tk->wall_to_monotonic;
1962 } while (read_seqcount_retry(&tk_core.seq, seq));
1964 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1965 now.tv_nsec + mono.tv_nsec);
1971 * Must hold jiffies_lock
1973 void do_timer(unsigned long ticks)
1975 jiffies_64 += ticks;
1976 calc_global_load(ticks);
1980 * ktime_get_update_offsets_now - hrtimer helper
1981 * @cwsseq: pointer to check and store the clock was set sequence number
1982 * @offs_real: pointer to storage for monotonic -> realtime offset
1983 * @offs_boot: pointer to storage for monotonic -> boottime offset
1984 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1986 * Returns current monotonic time and updates the offsets if the
1987 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
1990 * Called from hrtimer_interrupt() or retrigger_next_event()
1992 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
1993 ktime_t *offs_boot, ktime_t *offs_tai)
1995 struct timekeeper *tk = &tk_core.timekeeper;
2001 seq = read_seqcount_begin(&tk_core.seq);
2003 base = tk->tkr_mono.base;
2004 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2005 base = ktime_add_ns(base, nsecs);
2007 if (*cwsseq != tk->clock_was_set_seq) {
2008 *cwsseq = tk->clock_was_set_seq;
2009 *offs_real = tk->offs_real;
2010 *offs_boot = tk->offs_boot;
2011 *offs_tai = tk->offs_tai;
2014 /* Handle leapsecond insertion adjustments */
2015 if (unlikely(base.tv64 >= tk->next_leap_ktime.tv64))
2016 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2018 } while (read_seqcount_retry(&tk_core.seq, seq));
2024 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2026 int do_adjtimex(struct timex *txc)
2028 struct timekeeper *tk = &tk_core.timekeeper;
2029 unsigned long flags;
2030 struct timespec64 ts;
2034 /* Validate the data before disabling interrupts */
2035 ret = ntp_validate_timex(txc);
2039 if (txc->modes & ADJ_SETOFFSET) {
2040 struct timespec delta;
2041 delta.tv_sec = txc->time.tv_sec;
2042 delta.tv_nsec = txc->time.tv_usec;
2043 if (!(txc->modes & ADJ_NANO))
2044 delta.tv_nsec *= 1000;
2045 ret = timekeeping_inject_offset(&delta);
2050 getnstimeofday64(&ts);
2052 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2053 write_seqcount_begin(&tk_core.seq);
2055 orig_tai = tai = tk->tai_offset;
2056 ret = __do_adjtimex(txc, &ts, &tai);
2058 if (tai != orig_tai) {
2059 __timekeeping_set_tai_offset(tk, tai);
2060 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2062 tk_update_leap_state(tk);
2064 write_seqcount_end(&tk_core.seq);
2065 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2067 if (tai != orig_tai)
2070 ntp_notify_cmos_timer();
2075 #ifdef CONFIG_NTP_PPS
2077 * hardpps() - Accessor function to NTP __hardpps function
2079 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2081 unsigned long flags;
2083 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2084 write_seqcount_begin(&tk_core.seq);
2086 __hardpps(phase_ts, raw_ts);
2088 write_seqcount_end(&tk_core.seq);
2089 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2091 EXPORT_SYMBOL(hardpps);
2095 * xtime_update() - advances the timekeeping infrastructure
2096 * @ticks: number of ticks, that have elapsed since the last call.
2098 * Must be called with interrupts disabled.
2100 void xtime_update(unsigned long ticks)
2102 write_seqlock(&jiffies_lock);
2104 write_sequnlock(&jiffies_lock);